The Machine Behind the Money
How modern AI actually works
A plain-language account of what a large language model actually is, written for professionals who work with numbers rather than with code. It explains tokens, training and prediction; why generating text is different from looking something up; why fabricated output follows from the way these models work rather than from a fault that can be repaired; and it sets out the four kinds of AI the rest of the course relies on: narrow, generative, multimodal and agentic. By the end you will be able to judge, from how a tool works rather than from how it is described in a sales presentation, what it can and cannot be trusted with when money and personal data are involved.
- Explain in plain terms how a generative model produces text or an image, and why the same prompt can yield different and sometimes fabricated output.
- Distinguish narrow, generative, multimodal and agentic AI, and say which financial tasks each is suited to and unsuited to.
- Judge from how a tool works, rather than from how it is described by a supplier, what it can and cannot be trusted with when money and personal data are involved.
- Reading a model's claims critically
- Separating retrieval from generation
- Recognising hallucination and over-confidence
- Matching tool class to financial task
Lessons in this module
How it lands across the four desks
You learn why a synthetic document or a fabricated narrative can look flawless. The system that produced it was trained to generate material that resembles the genuine article, and it has no separate check on whether the content is true. That understanding of the mechanism underpins every detection skill taught in Modules 2 to 4.
You learn why a confident, fluent answer from an AI assistant can still be wrong, and measurably often is on legal questions, and why checking it against the primary source is a control your institution depends on rather than an optional extra.
You learn why a client-ready draft still needs checking, what the model does and does not know about your client, and how to use generative tools for speed without passing their fabrications on to the client.
You learn why a score is a prediction rather than a statement of fact. The same kind of statistical machinery that selects a chatbot's next word produces the probability behind a credit decision, and it carries the same limits.
Key literature · 6 sources
Every module rests on a verified scholarly and institutional evidence base. The full core and further reading lists open with the module.
- Vaswani, A. et al. (2017) 'Attention Is All You Need.' NeurIPS 2017. arxiv.org/abs/1706.03762: the transformer architecture underlying modern language models.
- Shanahan, M. (2024) 'Talking About Large Language Models.' Communications of the ACM 67(2): the discipline of describing what models actually do.
- Ji, Z. et al. (2023) 'Survey of Hallucination in Natural Language Generation.' ACM Computing Surveys 55(12): why fabrication is structural.
- Bender, E. M., Gebru, T., McMillan-Major, A. & Shmitchell, S. (2021) 'On the Dangers of Stochastic Parrots.' FAccT '21: pattern fluency without truth-tracking.
- Bommasani, R. et al. (2021) 'On the Opportunities and Risks of Foundation Models.' Stanford CRFM: why one model's failure modes propagate everywhere.
- SARB Prudential Authority & FSCA (2025) 'Artificial Intelligence in the South African Financial Sector': the local adoption, risk and skills picture; the sector-survey findings cited throughout this module.